The same 3D detector fine-tuned twice, once on the labelled scene in TIER IV's public t4-sample release as published and once on an edited version with the boxes that carry no lidar return removed. Same budget, same four class states, 180 trials each. The edited one scored 5.6 points higher.
A same-size random cut of 151 boxes landed inside the full arm's interval, so the gain comes from which boxes were removed and not from removing boxes.
The class state the cut hit hardest, standing pedestrians at 97 of 201 boxes gone, gains most. Parked cars come out flat at 180 trials. The pooled lift clears zero, single class states at this trial count mostly do not, and the chart says so.
Measured on every annotated box in the release, all 658 of them, no sampling, and num_radar_pts is zero on every one of them so a lidar return is the only sensor evidence a box carries. 151 hold none, and range does not explain that, it inverts it, since the median removed box sits 36.8 m from the lidar against 70.6 m for the boxes kept, and inside 20 m the rate is 65 of 148 against 21 of 350 beyond 60 m. 44 of the 96 boxes whose own label reads no occlusion of the object hold zero returns, at a median 5.6 m from the lidar, against 16.1 m for the fully visible boxes that do have returns. The same rate is in the 100 hand-labelled keyframe boxes the interpolated set was built from, 25 of 100 overall and 8 of the 15 fully visible ones, so this is a property of the labelling and not of the interpolation.
| Removed, boxes with no lidar return | boxes | rate |
|---|---|---|
| pedestrian standing | 97 | 48.3% |
| car parked | 53 | 13.2% |
| car driving | 1 | 3.0% |
| bicycle driving | 0 | 0% |
| Removed | 151 | 22.9% |
| Kept | 507 | 77.1% |